The building at 1585 Broadway does not look, from the outside, like a facility processing more than a billion words of financial text every twenty-four hours. Yet since January of this year, Morgan Stanley's internal document intelligence platform has done precisely that, consuming regulatory filings, sell-side research, earnings call transcripts, and internal memoranda at a rate that its designers describe, with some understatement, as a step change in the firm's analytical capacity.
The platform, which the bank has named Apex DI internally, is built atop a version of Meta's Llama 3 architecture fine-tuned on approximately four years of proprietary financial documents. What distinguishes it from earlier natural language processing experiments at the firm is not merely scale but precision: on structured extraction tasks, defined as the accurate identification and retrieval of specific data points from unstructured text, the system has achieved an accuracy rate of 97.6 per cent, a figure that the firm's own validation team acknowledges exceeds the performance of human analysts working under equivalent conditions.
What the Engine Does
The immediate applications are prosaic by the standards of recent artificial intelligence discourse: the system reads ten-K filings and extracts key financial metrics; it summarises earnings call transcripts for equity research analysts; it classifies regulatory correspondence by topic, urgency, and response obligation. These are tasks that, until recently, were performed by teams of junior analysts working under significant time pressure. The reassignment of that labour is, by most accounts within the firm, already well advanced.
The more consequential applications are less visible. Apex DI now sits at the beginning of the firm's investment idea pipeline, surfacing thematic connections across company disclosures that human analysts, operating within the time constraints of normal capital markets practice, would rarely have the bandwidth to identify. A comment buried in a mid-tier industrial company's annual report may, the system determines, carry implications for a supply chain relationship with a consumer technology company that appears elsewhere in the day's document queue. The system flags the connection. A human analyst decides whether it matters.
97.6% accuracy on structured extraction tasks surpasses what the firm's own validation team says human analysts achieve under equivalent conditions.
What It Cannot Do
The system's designers are, to their credit, more forthcoming about the platform's limitations than one might expect from an institution with a commercial interest in projecting technological confidence. Apex DI performs well on documents that conform to established structural patterns. It performs considerably less well on documents that deploy language in novel or deliberately ambiguous ways, which is to say, on precisely the documents most likely to contain material information that the market has not yet priced.
Legal disclaimers that have been crafted to satisfy a regulator while obscuring a material risk remain, the firm's own testing suggests, more reliably identified by experienced counsel than by the model. Financial statements that comply with accounting standards while presenting an optimistic picture of underlying economic reality present a similar challenge. The model has been trained on what the documents say; it has not been trained, and cannot easily be trained, on the gap between what documents say and what they mean.
This is not a criticism that Morgan Stanley's technology leadership would dispute. It is, rather, the basis on which they argue that Apex DI is a tool for augmenting analyst judgment rather than replacing it. The firm has been explicit, in internal communications reviewed by this publication, that no investment decision may be made on the basis of the platform's output alone. The constraint is sensible. Whether it will survive the commercial pressure to reduce headcount further, as the platform's capabilities expand, is a question that only time will answer.